sPAN-pROP: Combatting Contextualized Social Media State-Linked Propaganda in the Middle East

Social Cyber-attacks such as propaganda, conspiracy theories, anger, and hate discourse are very old phenomena that inflict harm to humans, organizations, national security, public officials, democratization efforts, careers, and policies. There have been significant efforts to identify anti-US speech on social media which includes propaganda. Such efforts largely ignore the attempts by other countries to manipulate social media in some regions including the Middle East. Research in this area is computational and solely focuses on fine tuning language models to detect general propaganda attacks. This paper addresses a new category of propaganda attacks that are tied to state-linked accounts that spread anti-US propaganda by taking advantages of specific geo-political crises in the Middle East. We investigated the role of general language models and training data to detect those forms of propaganda. Our study concludes that existing propaganda training data is unable to successfully detect targeted propaganda. We propose a contextualized span detection approach to identify these types of propaganda and show that our targeted training models work significantly better compared to the existing general propaganda detection approaches.